NPU and RKLLM support on Rockchip RK3588 (NanoPC-T6) and RK3576 (NanoPi M5)

[fever_wits]

Hello,

I am currently working on a project involving local LLM inference on Rockchip SoCs using the rknn-llm framework. I have two specific questions regarding Armbian support for the NPU:
 1. NanoPi M5 (RK3576): FriendlyElec recently released the NanoPi M5 based on the RK3576. If I use the Armbian image for this board, will I be able to run LLM models using the NPU? I am considering purchasing this board for LLM
    testing and would like to know if the driver support is already there.
 2. NanoPC-T6 (RK3588): I have a NanoPC-T6 running Armbian, but I cannot find the /dev/rknpu device node. I have tested this with both 6.1.115-vendor-rk35xx and 6.12.58-current-rockchip64 kernels, but the NPU device is not appearing.
    Is there a specific overlay or a manual configuration step needed to enable NPU support on these kernels?
 tried adding tags npu and nanopi m5, but there was no such option.
I am happy to provide any additional information, logs (such as `armbianmonitor -u`), or command outputs if needed to help diagnose the issue.
Thank you in advance for your help and guidance!

Best regards,
A.H.
 

[Werner]

7 hours ago, fever_wits said:

but I cannot find the /dev/rknpu device node

npu is there in both vendor bsp kernel 6.1.y. There is no node exposed afaik. Everthing beyond like libraries to use npu is userspace and out of scope of Armbian.

edge 6.18.y also has reverse-engineered npu support with a driver called Rocket.

[fever_wits]

Hello,
I think the problem with not seeing /dev/rknpu is that it is not defined in the rk3588-nanopc-t6.dtb file.
In earlier versions, there was no fan control and with each update I made corrections.
I have a saved dts file from 14.01.2025
 

206220 Dec 28 21:12 rk3588-nanopc-t6.dts
341802 Jan 14 2025 rk3588-nanopc-t6.dts.14.01.2025

There is a big difference between the two. I know that NanoPC-T6 is Community maintained, who can I contact to see what the problem is?
It is possible that I am also making a mistake in trying to run an LLM model.

Best regards,
A.H.